Homework correction method, system, readable medium and electronic device
By training a small sample character set model for each question and dynamically managing the model status, the problem of AI correction technology's low accuracy in handwritten character recognition is solved, and the accuracy of AI correction and the utilization rate of computing resources are improved.
Patent Information
- Application Number
- CN202210178875.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-02-25
- Publication Date
- 2024-12-20
- Estimated Expiration
- 2042-02-25
AI Technical Summary
The existing AI correction technology has low accuracy in identifying handwritten characters, which leads to relying on manual correction when correcting subjective questions and being unable to effectively replace manual correction.
Training a small sample character set model for each question, and dynamically manage the online and offline status of the model according to the target user's course schedule to improve the accuracy of AI correction and the utilization rate of computing resources.
The accuracy of AI correction has been improved, especially in the correction of subjective questions, the dependence on manual correction has been reduced, and computing resources have been effectively released.
Smart Images
Figure CN114581648B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular, to a homework grading method, system, readable medium and electronic device. Background Art
[0002] With the rapid development and rapid maturity of artificial intelligence technology and big data technology, the work of homework grading is gradually being replaced by AI (Artificial Intelligence) grading instead of manual grading. AI grading not only relieves the workload of teachers and parents, but also provides technical and data support and new business possibilities for subsequent adaptive learning.
[0003] However, the AI algorithm's recognition accuracy for handwritten characters cannot yet reach the same level of accuracy as printed text recognition. Therefore, the current business scenarios for AI automatic grading on the market are more focused on the single step of grading objective questions or simple math formulas (the character range and character changes involved in the questions and handwritten answers are relatively simple), while the grading of subjective questions still tends to use manual grading. Summary of the invention
[0004] The purpose of the present disclosure is to provide a homework grading method, system, readable medium and electronic device that can improve the accuracy of AI grading homework, so as to solve the problem that the accuracy of AI grading homework is not high.
[0005] In order to achieve the above object, the present disclosure provides a homework marking method, the method comprising:
[0006] Obtaining the identity of at least one target topic that needs to be corrected every day according to the target user's class schedule;
[0007] Pull up at least one corresponding target model in the model pool according to the identity of at least one target topic every day, and take offline the model whose service time has ended, wherein the model pool stores a plurality of models, each of which corresponds to a topic and the identity of the topic;
[0008] Receiving a homework picture sent by the target user, segmenting the homework picture to obtain at least one homework sub-picture, each of the homework sub-pictures including a question and an answer of the target user to the question;
[0009] According to the identity identifier of the topic included in each of the homework sub-pictures, forwarding each of the homework sub-pictures to the target model corresponding to the identity identifier for correction, to obtain the correction result of each of the homework sub-pictures;
[0010] After summarizing the correction results of each of the job sub-images, the correction result of the job image is obtained and returned to the target user.
[0011] Optionally, each model in the model pool is trained in the following manner:
[0012] For each of the models, a workflow is created, and multiple answer images of the questions corresponding to the model are imported into the workflow, each answer image including the question and the answer to the question;
[0013] Establishing the authority of annotators for each of the workflows, and receiving the annotation results of each of the answer pictures by authorized annotators, and using the annotation results of each of the answer pictures and the answer pictures as training samples, and adding them to the training set corresponding to the workflow;
[0014] The training set is trained to obtain the model.
[0015] Optionally, the method further comprises:
[0016] According to the identity of the topic included in each of the homework sub-pictures, each of the homework sub-pictures and the correction result of the homework sub-picture are imported into the corresponding workflow;
[0017] For the correction result of each of the job sub-images, the approval of the annotation personnel who have the authority to the workflow is received, and the correction result of each of the job sub-images after approval and the job sub-image are used as training samples and added to the training set corresponding to the workflow.
[0018] Optionally, the step of training the training set to obtain the model includes:
[0019] When the number of training samples in the training set reaches a first preset threshold, the preset model is not loaded, and the training set is fully trained to obtain the model; and / or,
[0020] When the number of newly added training samples in the training set reaches a second preset threshold, the model obtained in the previous training is loaded, and the newly added training samples in the training set are trained to obtain the model.
[0021] The present disclosure also provides a homework correction system, the system comprising:
[0022] A management module, used for obtaining the identity of at least one target topic that needs to be corrected every day according to the target user's class schedule;
[0023] A model deployment module, used to pull up at least one corresponding target model in the model pool according to the identity of the at least one target topic every day, and offline the model whose service time has ended, wherein the model pool stores a plurality of models, each of which corresponds to a topic and the identity of the topic;
[0024] a topic-cutting module, configured to receive the homework picture sent by the target user, and segment the homework picture to obtain at least one homework sub-picture, each of which includes a topic and the target user's answer to the topic;
[0025] A model correction module, used for forwarding each of the homework sub-images to a target model corresponding to the identity identifier for correction according to the identity identifier of the topic included in each of the homework sub-images, so as to obtain a correction result for each of the homework sub-images;
[0026] The summarizing module is used to summarize the correction results of each of the homework sub-images, obtain the correction results of the homework image and return them to the target user.
[0027] Optionally, the system further comprises:
[0028] A workflow creation module, for creating a workflow for each of the models, and importing multiple answer images of the questions corresponding to the model into the workflow, each answer image including the question and the answer to the question;
[0029] A labeling module, used to establish labeling personnel authority for each of the workflows, and receive labeling results of each of the answer pictures by authorized labeling personnel, and use the labeling results of each of the answer pictures and the answer pictures as training samples, and add them to the training set corresponding to the workflow;
[0030] The model training module is used to train the training set to obtain the model.
[0031] Optionally, the system further comprises:
[0032] A homework sub-image importing module, used for importing each homework sub-image and the correction result of the homework sub-image into a corresponding workflow according to the identity of the topic included in each homework sub-image;
[0033] The semi-automatic annotation module is used to receive the approval of the annotation personnel who have the authority to annotate the workflow for the correction results of each of the job sub-images, and use the approved correction results of each of the job sub-images and the job sub-images as training samples, and add them to the training set corresponding to the workflow.
[0034] Optionally, the model training module is specifically used for:
[0035] When the training samples in the training set reach a first preset threshold, the preset model is not loaded, and the training set is fully trained to obtain the model; and / or,
[0036] When the number of newly added training samples in the training set reaches a second preset threshold, the model obtained in the previous training is loaded, and the newly added training samples in the training set are trained to obtain the model.
[0037] The present disclosure also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.
[0038] The present disclosure also provides an electronic device, comprising:
[0039] a memory having a computer program stored thereon;
[0040] A processor is used to execute the computer program in the memory to implement the steps of the above method.
[0041] Through the above technical solution, each question corresponds to a model belonging to the question, which improves the accuracy of AI grading. In addition, according to the target user's class schedule, the online and offline status of each model is dynamically managed, the models that need to be used are started in advance, and the models that are not needed are offline, freeing up computing resources.
[0042] Other features and advantages of the present disclosure will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] The accompanying drawings are used to provide a further understanding of the present disclosure and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the present disclosure but do not constitute a limitation of the present disclosure. In the accompanying drawings:
[0044] Figure 1 It is a flow chart of a homework marking method provided by an embodiment of the present disclosure.
[0045] Figure 2 It is a block diagram of a homework grading system provided by an embodiment of the present disclosure.
[0046] Figure 3 It is a block diagram of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION
[0047] The specific implementation of the present disclosure is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described herein is only used to illustrate and explain the present disclosure, and is not used to limit the present disclosure.
[0048] In the process of implementing artificial intelligence algorithms in business scenarios, it is a recognized fact that when the scope of the business scenario is limited, the model trained with specific business information has higher accuracy and better performance than the general model.
[0049] Inspired by this, in order to further improve the accuracy of AI automated grading under the condition of limiting the scope of questions, we train a small sample character set model for each question, and the effect is much better than using a universal character set model. Take OCR recognition as an example: when it is known that the character set covered by a task is a specific set (such as 200 characters), the effect of training a model with a small character set is better than training a model with 6,000 common Chinese characters. For example, for the oral arithmetic grading task, if a model that only recognizes numbers + operation symbols is trained separately, the effect is much better than a model trained with a character set containing 6,000 Chinese characters + 10 numbers + 26×2 English letters.
[0050] Furthermore, considering that the actual call volume of these models for each topic is not evenly distributed, it depends on the advancement of the teacher's teaching tasks. Only after learning the knowledge point and assigning homework including the topic, the model corresponding to the topic will be called for correction. Therefore, in order to prevent the situation where many models will not be called for a long time but still occupy resources, the present invention dynamically manages the online and offline status of each model according to the target user's class schedule, starts the model that needs to be used in advance, and offlines the model that is not needed to release computing resources.
[0051] Based on the above inventive concept, an embodiment of the present disclosure provides a method for grading homework. Figure 1 FIG. 1 is a flowchart of a method for marking homework according to an embodiment of the present disclosure. Figure 1 As shown, the method comprises the following steps:
[0052] S10, obtaining the identity of at least one target topic that needs to be corrected every day according to the target user's class schedule.
[0053] The target user may refer to a user whose homework needs to be graded, and there may be multiple target users, such as students from multiple classes, etc. The target topic refers to the topic that the target user needs to grade every day.
[0054] S20, pulling up at least one corresponding target model in the model pool according to the identity identifier of the at least one target topic every day, and taking offline the model whose service time has ended.
[0055] The model pool stores multiple models, each of which corresponds to a topic and the identity of the topic. The target model is the model in the model pool that corresponds to the target topic. Through step S20, the model (target model) that is needed is put online in advance every day according to the target user's class schedule, and the model whose service time has ended (has been used up) is taken offline to release computing resources.
[0056] In a specific implementation, the model pool is mainly used to store model data and implement model management functions. It is responsible for recording each model identity, the data version used, the corresponding topic identity, accuracy indicators and other information. (1) Model and algorithm code splitting. Model and project code splitting is achieved through GIT and DVC solutions. (2) Model file management. Model version management and entity file management are achieved through the DVC solution. Each model file corresponds to a unique identity, so the model can be bundled with any project code, making dynamic publishing possible. (3) Model traceability. Through the mature industry technologies of flask / redis / sqlalchemy / celery based on Python, a distributed data collection solution is implemented, which realizes the collection and centralized display of training data under high concurrency.
[0057] A specific implementation method of step S20 can be performed by presetting a configuration file in advance, in which the model service that needs to be started and the time period that needs to be maintained can be configured, and regular patrols are performed. When it is found that the required model service in the configuration file is not started, the model service is automatically started. When the time period for model service maintenance ends, the process is automatically closed. Specifically, the code required to start the model service can be pre-set using code repositories such as gitlab, and the software environment for the model service to run can be configured using the cloud-native Harbor window image management platform. According to the hardware resources allocated by the management center (the management center can be set up so that the management center has the scheduling authority of global hardware resources, and can dynamically allocate computing resources for each service and integrate distributed computing power. The technical implementation of the management center can be completed using a container orchestration system based on Kubernetes), the model is automatically deployed in the form of a docker image.
[0058] S30, receiving the homework picture sent by the target user, segmenting the homework picture to obtain at least one homework sub-picture, each of the homework sub-pictures including a question and the target user's answer to the question.
[0059] Since the homework image may include multiple questions, and each model in this application corresponds to a question, it is necessary to segment the questions and divide the homework image into at least one homework sub-image, so that each of the homework sub-images includes a question and the target user's answer to the question.
[0060] S40, according to the identity identifier of the topic included in each of the homework sub-pictures, forwarding each of the homework sub-pictures to the target model corresponding to the identity identifier for correction, to obtain the correction result of each of the homework sub-pictures.
[0061] Each of the homework sub-images includes a topic. According to the identity identifier of the topic and the correspondence between the model and the identity identifier of the topic, each of the homework sub-images can be forwarded to the corresponding target model for correction to obtain the correction result of each of the homework sub-images.
[0062] Steps S30 and S40, by establishing a bus-forwarding mechanism, when a homework request that needs to be corrected comes in, the topic segmentation service is first performed. The topic identity (ID) that needs to be corrected can be obtained through the topic segmentation service, and the model identity corresponding to this topic identity is queried in the "Model and Topic Identity Correspondence Table". The bus service forwards the business request to the corresponding real model service, and the processing result is returned accordingly.
[0063] S50, after summarizing the correction results of each of the homework sub-pictures, the correction result of the homework picture is obtained and returned to the target user.
[0064] Through the above technical solution, each question corresponds to a model belonging to the question, which improves the accuracy of AI grading. In addition, according to the target user's class schedule, the online and offline status of each model is dynamically managed, the models that need to be used are started in advance, and the models that are not needed are offline, freeing up computing resources.
[0065] Optionally, each model in the model pool is trained in the following manner:
[0066] For each of the models, a workflow is created, and multiple answer images of the questions corresponding to the model are imported into the workflow, each answer image including the question and the answer to the question.
[0067] The authority of the annotator is established for each of the workflows, and the annotation results of each of the answer pictures by the authorized annotators are received. The annotation results of each of the answer pictures and the answer pictures are used as training samples and added to the training set corresponding to the workflow.
[0068] In a specific implementation, multiple (e.g., 10,000) answer pictures can be collected and organized for each question (sample question) and saved in a data center, and the question identity, campus, subject, grade, class, date, and batch of the answer pictures can be saved. Multiple workflows are created on the annotation platform (the number of workflows is the same as the number of questions, such as creating 1,000 workflows corresponding to 1,000 questions), and the (10,000) answer pictures belonging to each question are imported into the corresponding workflow, and the authority of the annotation personnel is established for each workflow to achieve authority isolation and provide protection for data security (if there is no authority restriction, all personnel involved in the annotation may leak data or models, and may also bring legal risks and hidden dangers to data involving user privacy).
[0069] Among them, in a specific implementation, the data center is mainly responsible for data management and encryption, and can configure permissions according to roles to ensure data security. It has the following functions: data version management; batch merging, splitting, and adding; permission management; data storage and screening. The data center is implemented using the industry's mature and stable Spring solution. Through DI and AOP in the Spring framework, data permissions are uniformly managed. The persistence of business data uses relational object storage and document-based distributed databases. The ORM layer uses advanced type-constrained jOOQ. File storage uses distributed file storage, which not only supports high data availability, but also ensures data security. Distributed file storage makes it possible to upload, convert, and download massive file data.
[0070] Among them, in a specific implementation, the annotation platform can use data specifications based on the PAGE-XML format to implement the data annotation-review-quality inspection-acceptance process through the workflow node method. The annotation platform can be composed of the following parts: (1) Front-end. The front-end is mainly responsible for specific annotation implementation, human-computer interaction, data display and other functions. The front-end of the annotation platform uses a complete G6 graph visualization engine based on relational data, providing an elegantly designed and easy-to-use workflow visualization solution, which not only ensures the flexibility and high scalability of the annotation template, but also realizes customized configuration at the node level granularity. At the same time, the stability of the front-end of the annotation platform ensures that the data is processed in an orderly and reliable manner in the workflow. In OCR annotation, a simple and practical third-party annotation tool is encapsulated, and its functions are customized for business scenarios based on its high scalability, realizing a highly usable visual OCR annotation tool that can meet the data annotation needs of the entire process. In addition to the visual OCR annotation tool, the front-end of the annotation platform also combines a visual ASR annotation tool, a data set import and export tool, etc., to provide a one-stop visual annotation solution with high usability, high reliability and high flexibility. (2) Back-end. The stable and reliable backend ensures the data consistency of parallel workflows. Through the low-latency, high-concurrency, high-availability, and highly reliable distributed message middleware, it smooths out the peaks and fills the valleys, ensuring that every business message can be delivered and processed. In the application server, the distributed task scheduling framework is innovatively used to inspect each piece of business data. By using the microservice design architecture, the backend annotation system can be expanded horizontally, which not only greatly improves the system throughput, but also ensures the availability of the service. In the architecture of the backend of the annotation system, in addition to the application server, it also combines message consumers and task executors. Each part of the service performs its own duties and together forms a highly available, high-performance, and high-concurrency annotation platform.
[0071] The training set is trained to obtain the model.
[0072] Through the above technical solution, a workflow is created for each question, and the authority of the annotation personnel is established for each workflow to achieve authority isolation and provide protection for data security. The multiple answer images corresponding to each question are imported into the corresponding workflow for annotation, and the training set of each question is obtained. By training the training set of each question, the model corresponding to each question can be obtained.
[0073] For each question, after obtaining the model corresponding to the question, the model can be applied to correct the homework sub-image including the question, and the correction result of the homework sub-image can be obtained. After that, the homework sub-image and the correction result of the homework sub-image can be saved, such as pushed to the data center as newly added data to be annotated. After the annotator manually corrects the correction result of the homework sub-image on the annotation platform, the corrected correction result of the homework sub-image and the homework sub-image are used as training samples and added to the training set corresponding to the workflow, thereby realizing semi-automatic annotation of the homework sub-image.
[0074] That is, optionally, the method further comprises:
[0075] According to the identity identifier of the topic included in each of the homework sub-pictures, each of the homework sub-pictures and the correction result of the homework sub-picture are imported into the corresponding workflow.
[0076] For the correction result of each of the job sub-images, the approval of the annotation personnel who have the authority to the workflow is received, and the correction result of each of the job sub-images after approval and the job sub-image are used as training samples and added to the training set corresponding to the workflow.
[0077] Approval means review and correction. Approval of the correction results of the homework sub-images can change the correction results or not. Therefore, the training samples in the training set corresponding to each question can include multiple manually annotated answer images, and can also include multiple homework sub-images that have been approved (semi-automatically annotated) by the model correction results.
[0078] Through the above technical solution, manual review is performed on the homework sub-images obtained after model correction, semi-automatic labeling of the homework sub-images is achieved, and the training set is expanded.
[0079] Optionally, the step of training the training set to obtain the model includes:
[0080] When the number of training samples in the training set reaches a first preset threshold, the preset model is not loaded, and the training set is fully trained to obtain the model.
[0081] And / or, when the number of newly added training samples in the training set reaches a second preset threshold, the model obtained in the previous training is loaded, and the newly added training samples in the training set are trained to obtain the model.
[0082] Initially, full training can be performed, that is, when the training samples in the training set reach a first preset threshold, the preset model is not loaded, and the training set is fully trained to obtain the model. After a model is obtained, as the training samples in the training set increase, incremental training can be performed, that is, when the newly added training samples in the training set reach a second preset threshold, the model obtained from the previous training is loaded, and the newly added training samples in the training set are trained to obtain the model.
[0083] In a specific embodiment, the current number of training samples can be read regularly at a fixed time period, and the model training is started after the current number of training samples reaches a first preset threshold (such as 10,000) or the newly added training samples reach a second preset threshold (such as 5,000). After starting the model training, the system pulls the training code from the code warehouse (for example, the algorithm training code is pre-configured using a code management warehouse such as Gitlab, and the corresponding warehouse is automatically pulled for model training when the training is started), and reads the set training mode. When it is a full training mode, there is no need to load the pre-training code to train directly. When the training mode is incremental training, the model of the previous training is obtained from the model pool as a pre-training model and loaded into the network. Start training. Generally, the same batch of training samples will be repeatedly trained for several rounds each time. After each round, the pre-prepared test data is automatically used for indicator evaluation. When the evaluation index does not reach the set threshold, continue training and repeat the above operations. When the evaluation index reaches the set threshold, stop training. The current model and training related information (such as the training sample batches used, training time, model identity, corresponding question identity, evaluation indicators, etc.) are saved in the model pool.
[0084] In the above implementation, the entire model training and termination are completed automatically by the script, without human intervention. The engineer's workload is just to prepare the training code and test data in advance, and provide the start and termination indicators, and this part of the work only needs to be performed once, without repeated operations.
[0085] It can be seen from the above that the homework grading method provided by the embodiments of the present disclosure can provide one-stop sample labeling, model training, deployment and application services.
[0086] Based on the above inventive concept, the disclosed embodiment also provides a homework grading system. Figure 2 FIG. 1 is a block diagram of a homework grading system according to an embodiment of the present disclosure. Figure 2 As shown, the homework marking system includes:
[0087] A management module, used for obtaining the identity of at least one target topic that needs to be corrected every day according to the target user's class schedule;
[0088] A model deployment module, used to pull up at least one corresponding target model in the model pool according to the identity of the at least one target topic every day, and offline the model whose service time has ended, wherein the model pool stores a plurality of models, each of which corresponds to a topic and the identity of the topic;
[0089] a topic-cutting module, configured to receive the homework picture sent by the target user, and segment the homework picture to obtain at least one homework sub-picture, each of which includes a topic and the target user's answer to the topic;
[0090] A model correction module, used for forwarding each of the homework sub-images to a target model corresponding to the identity identifier for correction according to the identity identifier of the topic included in each of the homework sub-images, so as to obtain a correction result for each of the homework sub-images;
[0091] The summarizing module is used to summarize the correction results of each of the homework sub-images, obtain the correction results of the homework image and return them to the target user.
[0092] Through the above technical solution, each question corresponds to a model belonging to the question, which improves the accuracy of AI grading. In addition, according to the target user's class schedule, the online and offline status of each model is dynamically managed, the models that need to be used are started in advance, and the models that are not needed are offline, freeing up computing resources.
[0093] Optionally, the system further comprises:
[0094] A workflow creation module, for creating a workflow for each of the models, and importing multiple answer images of the questions corresponding to the model into the workflow, each answer image including the question and the answer to the question;
[0095] A labeling module, used to establish labeling personnel authority for each of the workflows, and receive labeling results of each of the answer pictures by authorized labeling personnel, and use the labeling results of each of the answer pictures and the answer pictures as training samples, and add them to the training set corresponding to the workflow;
[0096] The model training module is used to train the training set to obtain the model.
[0097] Through the above technical solution, a workflow is created for each question, and the authority of the annotation personnel is established for each workflow to achieve authority isolation and provide protection for data security. The multiple answer images corresponding to each question are imported into the corresponding workflow for annotation, and the training set of each question is obtained. By training the training set of each question, the model corresponding to each question can be obtained.
[0098] Optionally, the system further comprises:
[0099] A homework sub-image importing module, used for importing each homework sub-image and the correction result of the homework sub-image into a corresponding workflow according to the identity of the topic included in each homework sub-image;
[0100] The semi-automatic annotation module is used to receive the approval of the annotation personnel who have the authority to annotate the workflow for the correction results of each of the job sub-images, and use the approved correction results of each of the job sub-images and the job sub-images as training samples, and add them to the training set corresponding to the workflow.
[0101] Through the above technical solution, manual review is performed on the homework sub-images obtained after model correction, semi-automatic labeling of the homework sub-images is achieved, and the training set is expanded.
[0102] Optionally, the model training module is specifically used for:
[0103] When the training samples in the training set reach a first preset threshold, the preset model is not loaded, and the training set is fully trained to obtain the model; and / or,
[0104] When the number of newly added training samples in the training set reaches a second preset threshold, the model obtained in the previous training is loaded, and the newly added training samples in the training set are trained to obtain the model.
[0105] Regarding the system in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.
[0106] Figure 3 FIG. 7 is a block diagram of an electronic device 700 according to an exemplary embodiment. Figure 3 As shown, the electronic device 700 may include: a processor 701 and a memory 702. The electronic device 700 may also include one or more of a multimedia component 703, an input / output (I / O) interface 704, and a communication component 705.
[0107] The processor 701 is used to control the overall operation of the electronic device 700 to complete all or part of the steps in the above-mentioned homework correction method. The memory 702 is used to store various types of data to support the operation of the electronic device 700. For example, these data may include instructions for any application or method used to operate on the electronic device 700, and application-related data, such as contact data, sent and received messages, pictures, audio, video, etc. The memory 702 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, referred to as EPROM), programmable read-only memory (Programmable Read-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic memory, flash memory, magnetic disk or optical disk. The multimedia component 703 may include a screen and an audio component. The screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signal may be further stored in the memory 702 or sent through the communication component 705. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 704 provides an interface between the processor 701 and other interface modules, and the other interface modules may be keyboards, mice, buttons, etc. These buttons may be virtual buttons or physical buttons. The communication component 705 is used for wired or wireless communication between the electronic device 700 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, 4G, NB-IOT, eMTC, or other 5G, etc., or a combination of one or more of them, is not limited here. Therefore, the corresponding communication component 705 may include: Wi-Fi module, Bluetooth module, NFC module, etc.
[0108] In an exemplary embodiment, the electronic device 700 can be implemented by one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), controllers, microcontrollers, microprocessors or other electronic components to execute the above-mentioned homework grading method.
[0109] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, and when the program instructions are executed by a processor, the steps of the above-mentioned homework correction method are implemented. For example, the computer-readable storage medium can be the above-mentioned memory 702 including program instructions, and the above-mentioned program instructions can be executed by the processor 701 of the electronic device 700 to complete the above-mentioned homework correction method.
[0110] In another exemplary embodiment, a computer program product is further provided. The computer program product includes a computer program that can be executed by a programmable device. The computer program has a code portion for executing the above-mentioned homework marking method when executed by the programmable device.
[0111] The preferred embodiments of the present disclosure are described in detail above in conjunction with the accompanying drawings; however, the present disclosure is not limited to the specific details in the above embodiments. Within the technical concept of the present disclosure, a variety of simple modifications can be made to the technical solution of the present disclosure, and these simple modifications all fall within the protection scope of the present disclosure.
[0112] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present disclosure will not further describe various possible combinations.
[0113] In addition, various embodiments of the present disclosure may be arbitrarily combined, and as long as they do not violate the concept of the present disclosure, they should also be regarded as the contents disclosed by the present disclosure.
Claims
1. A method for marking homework, It is characterized in that The method comprises: Obtaining the identity of at least one target topic that needs to be corrected every day according to the target user's class schedule; Pull up at least one corresponding target model in the model pool according to the identity of at least one target topic every day, and take offline the model whose service time has ended, wherein the model pool stores a plurality of models, each of which corresponds to a topic and the identity of the topic; Receiving a homework picture sent by the target user, segmenting the homework picture to obtain at least one homework sub-picture, each of the homework sub-pictures including a question and an answer of the target user to the question; According to the identity identifier of the topic included in each of the homework sub-pictures, forwarding each of the homework sub-pictures to the target model corresponding to the identity identifier for correction, to obtain the correction result of each of the homework sub-pictures; After summarizing the correction results of each of the homework sub-pictures, the correction results of the homework picture are obtained and returned to the target user; Each model in the model pool is trained in the following way: For each of the models, a workflow is created, and multiple answer images of the questions corresponding to the model are imported into the workflow, each answer image including the question and the answer to the question; Establishing the authority of annotators for each of the workflows, and receiving the annotation results of each of the answer pictures by authorized annotators, and using the annotation results of each of the answer pictures and the answer pictures as training samples, and adding them to the training set corresponding to the workflow; Training the training set to obtain the model; The method further comprises: According to the identity of the topic included in each of the homework sub-pictures, each of the homework sub-pictures and the correction result of the homework sub-picture are imported into the corresponding workflow; For the correction result of each of the job sub-images, the approval of the annotation personnel who have the authority to the workflow is received, and the correction result of each of the job sub-images after approval and the job sub-image are used as training samples and added to the training set corresponding to the workflow.
2. The method according to claim 1, It is characterized in that The step of training the training set to obtain the model comprises: When the number of training samples in the training set reaches a first preset threshold, the preset model is not loaded, and the training set is fully trained to obtain the model; and / or, When the number of newly added training samples in the training set reaches a second preset threshold, the model obtained in the previous training is loaded, and the newly added training samples in the training set are trained to obtain the model.
3. A homework grading system, It is characterized in that The system comprises: A management module, used for obtaining the identity of at least one target topic that needs to be corrected every day according to the target user's class schedule; A model deployment module, used to pull up at least one corresponding target model in the model pool according to the identity of the at least one target topic every day, and offline the model whose service time has ended, wherein the model pool stores a plurality of models, each of which corresponds to a topic and the identity of the topic; a topic-cutting module, configured to receive the homework picture sent by the target user, and segment the homework picture to obtain at least one homework sub-picture, each of which includes a topic and the target user's answer to the topic; A model correction module, used for forwarding each of the homework sub-images to a target model corresponding to the identity identifier for correction according to the identity identifier of the topic included in each of the homework sub-images, so as to obtain a correction result for each of the homework sub-images; A summarizing module, used to summarize the correction results of each of the homework sub-pictures, obtain the correction results of the homework picture and return them to the target user; Wherein, the system further comprises: A workflow creation module, for creating a workflow for each of the models, and importing multiple answer images of the questions corresponding to the model into the workflow, each answer image including the question and the answer to the question; A labeling module, used to establish labeling personnel authority for each of the workflows, and receive labeling results of each of the answer pictures by authorized labeling personnel, and use the labeling results of each of the answer pictures and the answer pictures as training samples, and add them to the training set corresponding to the workflow; A model training module, used to train the training set to obtain the model; The system further comprises: A homework sub-image importing module, used for importing each homework sub-image and the correction result of the homework sub-image into a corresponding workflow according to the identity of the topic included in each homework sub-image; The semi-automatic annotation module is used to receive the approval of the annotation personnel who have the authority to annotate the workflow for the correction results of each of the job sub-images, and use the approved correction results of each of the job sub-images and the job sub-images as training samples, and add them to the training set corresponding to the workflow.
4. The system according to claim 3, It is characterized in that The model training module is specifically used for: When the training samples in the training set reach a first preset threshold, the preset model is not loaded, and the training set is fully trained to obtain the model; and / or, When the number of newly added training samples in the training set reaches a second preset threshold, the model obtained in the previous training is loaded, and the newly added training samples in the training set are trained to obtain the model.
5. A non-transitory computer-readable storage medium having a computer program stored thereon, It is characterized in that When the program is executed by a processor, the steps of the method according to claim 1 or 2 are implemented.
6. An electronic device, It is characterized in that include: a memory having a computer program stored thereon; A processor, configured to execute the computer program in the memory to implement the steps of the method according to claim 1 or 2.